COMPMID-2110: Enable CLGEMMLowpMatrixMultiplyReshapeOnlyRHSKernel in CLGEMMLowp

Change-Id: Ic32c803c3e2a067de10a7e46c85c962a970957b6
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/969
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 17 insertions, 44 deletions

diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index c447cb8778..cd537087d3 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -24,7 +24,7 @@
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
 
 #include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h"
+#include "arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/TensorInfo.h"
@@ -40,17 +40,16 @@ using namespace arm_compute::cl_gemm;
 
 namespace
 {
-inline bool is_gemm_reshaped(unsigned int m, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
+inline bool is_gemm_reshaped(bool reshape_b_only_on_first_run, GPUTarget gpu_target)
 {
-    return (get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) && (m > 1) && (reshape_b_only_on_first_run);
+    return (get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) && (reshape_b_only_on_first_run);
 }
 } // namespace
 
 CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)),
       _mm_kernel(),
-      _mm_reshaped_kernel(),
-      _mtx_a_reshape_kernel(),
+      _mm_reshaped_only_rhs_kernel(),
       _mtx_b_reshape_kernel(),
       _mtx_a_reduction_kernel(),
       _mtx_b_reduction_kernel(),
@@ -58,7 +57,6 @@ CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
       _offset_contribution_output_stage_kernel(),
       _vector_sum_col(),
       _vector_sum_row(),
-      _tmp_a(),
       _tmp_b(),
       _mm_result_s32(),
       _original_b(nullptr),
@@ -86,7 +84,6 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     const GPUTarget gpu_target = CLScheduler::get().target();
 
     // Set the target for the kernels
-    _mtx_a_reshape_kernel.set_target(gpu_target);
     _mm_kernel.set_target(gpu_target);
 
     const ICLTensor *matrix_a = a;
@@ -105,27 +102,19 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
 
     // Check if we need to reshape the matrix A and matrix B
-    _is_gemm_reshaped = is_gemm_reshaped(m, _reshape_b_only_on_first_run, gpu_target);
+    _is_gemm_reshaped = is_gemm_reshaped(_reshape_b_only_on_first_run, gpu_target);
 
     if(_is_gemm_reshaped)
     {
-        // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
-        reinterpret_input_as_3d = false;
-
-        matrix_a = &_tmp_a;
         matrix_b = &_tmp_b;
 
-        _memory_group.manage(&_tmp_a);
         if(!_reshape_b_only_on_first_run)
         {
             _memory_group.manage(&_tmp_b);
         }
 
         // Pick up the GEMM configuration
-        std::tie(lhs_info, rhs_info) = CLGEMMReshapedKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
-        // Configure reshape LHS kernel
-        _mtx_a_reshape_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
+        std::tie(lhs_info, rhs_info) = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
 
         // Configure reshape RHS kernel
         _mtx_b_reshape_kernel.configure(b, &_tmp_b, rhs_info);
@@ -166,7 +155,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         if(_is_gemm_reshaped)
         {
             // Configure and tune matrix multiply kernel
-            _mm_reshaped_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+            _mm_reshaped_only_rhs_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
         }
         else
         {
@@ -185,7 +174,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         if(_is_gemm_reshaped)
         {
             // Configure and tune matrix multiply kernel
-            _mm_reshaped_kernel.configure(matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+            _mm_reshaped_only_rhs_kernel.configure(matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
         }
         else
         {
@@ -200,7 +189,6 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     // Allocate tensors
     if(_is_gemm_reshaped)
     {
-        _tmp_a.allocator()->allocate();
         if(!_reshape_b_only_on_first_run)
         {
             _tmp_b.allocator()->allocate();
@@ -231,7 +219,6 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     const ITensorInfo *matrix_a_info = a;
     const ITensorInfo *matrix_b_info = b;
 
-    TensorInfo        tmp_a_info{};
     TensorInfo        tmp_b_info{};
     GEMMRHSMatrixInfo rhs_info;
     GEMMLHSMatrixInfo lhs_info;
@@ -246,27 +233,16 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     const unsigned int batch_size              = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
     const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
 
-    bool reshape_matrices = is_gemm_reshaped(m, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
-
-    // if reshape_matrices is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
-    if(reshape_matrices)
-    {
-        reinterpret_input_as_3d = false;
-    }
+    bool reshape_matrix_b = is_gemm_reshaped(gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
 
     const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d);
 
-    if(reshape_matrices)
+    if(reshape_matrix_b)
     {
-        matrix_a_info = &tmp_a_info;
         matrix_b_info = &tmp_b_info;
 
         // Pick up the GEMM configuration
-        std::tie(lhs_info, rhs_info) = CLGEMMReshapedKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
-        // Validate reshape LHS kernel
-        auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d())));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeLHSMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d()));
+        std::tie(lhs_info, rhs_info) = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
 
         // Validate reshape RHS kernel
         auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
@@ -297,13 +273,13 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     {
         TensorInfo mm_result_s32_info{};
 
-        if(reshape_matrices)
+        if(reshape_matrix_b)
         {
             // Output tensor auto inizialitation if not yet initialized
             auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32));
 
             // Validate matrix multiply
-            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
         }
         else
         {
@@ -324,10 +300,10 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     }
     else
     {
-        if(reshape_matrices)
+        if(reshape_matrix_b)
         {
             // Validate matrix multiply
-            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
         }
         else
         {
@@ -356,9 +332,6 @@ void CLGEMMLowpMatrixMultiplyCore::run()
 
     if(_is_gemm_reshaped)
     {
-        // Run reshape matrix A
-        CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false);
-
         if(!_reshape_b_only_on_first_run)
         {
             // Run reshape matrix B
@@ -375,7 +348,7 @@ void CLGEMMLowpMatrixMultiplyCore::run()
     // Run matrix multiply
     if(_is_gemm_reshaped)
     {
-        CLScheduler::get().enqueue(_mm_reshaped_kernel, false);
+        CLScheduler::get().enqueue(_mm_reshaped_only_rhs_kernel, false);
     }
     else
     {
@@ -425,4 +398,4 @@ void CLGEMMLowpMatrixMultiplyCore::prepare()
         _is_prepared = true;
     }
 }
-} // namespace arm_compute
+} // namespace arm_compute
\ No newline at end of file